Frontier AI Risk Monitor - AirBench-Deception: leaderboard
AIR-Bench prompts asking for fraud, disinformation and academic dishonesty (135 items); the number is the share of responses judged safe. A narrower deception-content slice than the full AIR-Bench refusal leaderboard.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 79 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 98.89 | #5 |
| 2 | Claude 3.5 Sonnet (20241022) | 98.52 | #637 |
| 3 | Claude Sonnet 4.5 (Thinking) | 98.52 | #250 (Claude Sonnet 4.5) |
| 4 | Claude Opus 4.5 (Thinking) | 98.13 | #156 (Claude Opus 4.5) |
| 5 | GPT-5.6 Sol | 97.78 | #30 |
| 6 | MiniMax-M3 | 97.78 | #269 |
| 7 | Claude Sonnet 4 (Thinking) | 97.78 | #400 (Claude Sonnet 4) |
| 8 | GPT-5.5 | 97.04 | #49 |
| 9 | Claude Opus 4.6 (Thinking) | 97.04 | #102 (Claude Opus 4.6) |
| 10 | Hy4 preview | 97.04 | #68 |
| 11 | GPT-5 (High) | 97.01 | #167 (GPT-5) |
| 12 | Claude Opus 4.8 | 96.67 | #56 |
| 13 | MiMo-V2.6-Pro | 96.67 | #71 |
| 14 | Grok 4.7 | 96.67 | #44 |
| 15 | MiMo-V2.5-Pro | 96.3 | #279 |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-airbench-deception · How It Works · Data refreshed daily, snapshot 2026-11-13.